Emotion-based music recommendation using audio features and user playlist

In this paper we utilize a dimensional emotion representation named Resonance-Arousal-Valence to express music emotion and inverse exponential function to represent emotion decay process. The relationship between acoustic features and their emotional impact reflection based on this representation has been well constructed. As music well expresses feelings, through the users' historical playlist in a session, we utilize the Conditional Random Fields to compute the probabilities of different emotion states, choosing the largest as the predicted user's emotion state. In order to recommend music based on the predicted user's emotion, we choose the optimized ranked music list that has the highest emotional similarities to the music invoking the predicted emotion state in the playlist for recommendation. We utilize our minimization iteration algorithm to assemble the optimized ranked recommended music list. The experiment results show that the proposed emotion-based music recommendation paradigm is effective to track the user's emotions and recommend music fitting his emotional state.